Nikola Milošević

Adela Ljajić, Miloš Košprdić, Bojana Bašaragin et al.

In this paper, we introduce the VerifAI project, a pioneering open-source scientific question-answering system, designed to provide answers that are not only referenced but also automatically vetted and verifiable. The components of the system are (1) an Information Retrieval system combining semantic and lexical search techniques over scientific papers (PubMed), (2) a Retrieval-Augmented Generation (RAG) module using fine-tuned generative model (Mistral 7B) and retrieved articles to generate claims with references to the articles from which it was derived, and (3) a Verification engine, based on a fine-tuned DeBERTa and XLM-RoBERTa models on Natural Language Inference task using SciFACT dataset. The verification engine cross-checks the generated claim and the article from which the claim was derived, verifying whether there may have been any hallucinations in generating the claim. By leveraging the Information Retrieval and RAG modules, Verif.ai excels in generating factual information from a vast array of scientific sources. At the same time, the Verification engine rigorously double-checks this output, ensuring its accuracy and reliability. This dual-stage process plays a crucial role in acquiring and confirming factual information, significantly enhancing the information landscape. Our methodology could significantly enhance scientists' productivity, concurrently fostering trust in applying generative language models within scientific domains, where hallucinations and misinformation are unacceptable.

Nikola Milosevic

As Large Language Models (LLMs) evolve into persistent scientific collaborators, context window saturation has emerged as a critical bottleneck. Scientific workflows involving iterative data analysis and hypothesis refinement rapidly saturate even extended contexts with dense technical content, while monolithic approaches suffer from quadratic cost scaling and cognitive degradation. We evaluate a Dual Process Memory Architecture that decouples immediate episodic needs (constant 10-message window) from long-term consolidated knowledge (growing at approximately 3 tokens/message). Unlike prior social agent memory systems, our domain-specific consolidation addresses contradictory parameter evolution, multi-hop reasoning across experimental phases, and precise technical fact retention. Through large-scale evaluation spanning 15,000 messages with cross-model validation across six LLMs from three families (OpenAI, Anthropic, Google), totaling 1,440 queries, we establish three key findings. First, while full-context models fail at 10,000 messages due to context overflow, our system maintains 70-85% accuracy with 1-2 second latency using 62% fewer tokens (45,434 vs 120,000+ limit). Second, cross-model validation reveals architecture-level trade-offs independent of specific LLMs: Dual Process excels at numeric/temporal queries (65-90% accuracy) while RAG excels at historical retrieval (60-85%), suggesting complementary deployment strategies. Third, we identify a "Sim-to-Real" gap where synthetic tests maintain constant memory but realistic workflows exhibit linear growth (about 3 tokens/message), with consolidation quality emerging as the primary scalability bottleneck. The architecture successfully manages profiles with 14,000+ scientific facts (125k tokens), demonstrating that domain-specific memory consolidation enables sustained operation beyond full-context limits.

Miloš Košprdić, Adela Ljajić, Bojana Bašaragin et al.

We introduce VerifAI, an open-source expert system for biomedical question answering that integrates retrieval-augmented generation (RAG) with a novel post-hoc claim verification mechanism. Unlike standard RAG systems, VerifAI ensures factual consistency by decomposing generated answers into atomic claims and validating them against retrieved evidence using a fine-tuned natural language inference (NLI) engine. The system comprises three modular components: (1) a hybrid Information Retrieval (IR) module optimized for biomedical queries (MAP@10 of 42.7%), (2) a citation-aware Generative Component fine-tuned on a custom dataset to produce referenced answers, and (3) a Verification Component that detects hallucinations with state-of-the-art accuracy, outperforming GPT-4 on the HealthVer benchmark. Evaluations demonstrate that VerifAI significantly reduces hallucinated citations compared to zero-shot baselines and provides a transparent, verifiable lineage for every claim. The full pipeline, including code, models, and datasets, is open-sourced to facilitate reliable AI deployment in high-stakes domains.

Aleksandar Kovačević, Bojana Bašaragin, Nikola Milošević et al.

Background: Electronic health records (EHRs) are a valuable resource for data-driven medical research. However, the presence of protected health information (PHI) makes EHRs unsuitable to be shared for research purposes. De-identification, i.e. the process of removing PHI is a critical step in making EHR data accessible. Natural language processing has repeatedly demonstrated its feasibility in automating the de-identification process. Objectives: Our study aims to provide systematic evidence on how the de-identification of clinical free text has evolved in the last thirteen years, and to report on the performances and limitations of the current state-of-the-art systems. In addition, we aim to identify challenges and potential research opportunities in this field. Methods: A systematic search in PubMed, Web of Science and the DBLP was conducted for studies published between January 2010 and February 2023. Titles and abstracts were examined to identify the relevant studies. Selected studies were then analysed in-depth, and information was collected on de-identification methodologies, data sources, and measured performance. Results: A total of 2125 publications were identified for the title and abstract screening. 69 studies were found to be relevant. Machine learning (37 studies) and hybrid (26 studies) approaches are predominant, while six studies relied only on rules. Majority of the approaches were trained and evaluated on public corpora. The 2014 i2b2/UTHealth corpus is the most frequently used (36 studies), followed by the 2006 i2b2 (18 studies) and 2016 CEGS N-GRID (10 studies) corpora.

Daniel Steinigen, Roman Teucher, Timm Heine Ruland et al.

Recent advancements in Large Language Models (LLMs) have showcased their proficiency in answering natural language queries. However, their effectiveness is hindered by limited domain-specific knowledge, raising concerns about the reliability of their responses. We introduce a hybrid system that augments LLMs with domain-specific knowledge graphs (KGs), thereby aiming to enhance factual correctness using a KG-based retrieval approach. We focus on a medical KG to demonstrate our methodology, which includes (1) pre-processing, (2) Cypher query generation, (3) Cypher query processing, (4) KG retrieval, and (5) LLM-enhanced response generation. We evaluate our system on a curated dataset of 69 samples, achieving a precision of 78\% in retrieving correct KG nodes. Our findings indicate that the hybrid system surpasses a standalone LLM in accuracy and completeness, as verified by an LLM-as-a-Judge evaluation method. This positions the system as a promising tool for applications that demand factual correctness and completeness, such as target identification -- a critical process in pinpointing biological entities for disease treatment or crop enhancement. Moreover, its intuitive search interface and ability to provide accurate responses within seconds make it well-suited for time-sensitive, precision-focused research contexts. We publish the source code together with the dataset and the prompt templates used.

Bojana Bašaragin, Adela Ljajić, Darija Medvecki et al.

Large language models (LLMs) have recently become the leading source of answers for users' questions online. Despite their ability to offer eloquent answers, their accuracy and reliability can pose a significant challenge. This is especially true for sensitive domains such as biomedicine, where there is a higher need for factually correct answers. This paper introduces a biomedical retrieval-augmented generation (RAG) system designed to enhance the reliability of generated responses. The system is based on a fine-tuned LLM for the referenced question-answering, where retrieved relevant abstracts from PubMed are passed to LLM's context as input through a prompt. Its output is an answer based on PubMed abstracts, where each statement is referenced accordingly, allowing the users to verify the answer. Our retrieval system achieves an absolute improvement of 23% compared to the PubMed search engine. Based on the manual evaluation on a small sample, our fine-tuned LLM component achieves comparable results to GPT-4 Turbo in referencing relevant abstracts. We make the dataset used to fine-tune the models and the fine-tuned models based on Mistral-7B-instruct-v0.1 and v0.2 publicly available.

Miloš Košprdić, Adela Ljajić, Bojana Bašaragin et al.

In this paper, we present the current progress of the project Verif.ai, an open-source scientific generative question-answering system with referenced and verified answers. The components of the system are (1) an information retrieval system combining semantic and lexical search techniques over scientific papers (PubMed), (2) a fine-tuned generative model (Mistral 7B) taking top answers and generating answers with references to the papers from which the claim was derived, and (3) a verification engine that cross-checks the generated claim and the abstract or paper from which the claim was derived, verifying whether there may have been any hallucinations in generating the claim. We are reinforcing the generative model by providing the abstract in context, but in addition, an independent set of methods and models are verifying the answer and checking for hallucinations. Therefore, we believe that by using our method, we can make scientists more productive, while building trust in the use of generative language models in scientific environments, where hallucinations and misinformation cannot be tolerated.

Darija Medvecki, Bojana Bašaragin, Adela Ljajić et al.

This paper presents the results of the first application of BERTopic, a state-of-the-art topic modeling technique, to short text written in a morphologi-cally rich language. We applied BERTopic with three multilingual embed-ding models on two levels of text preprocessing (partial and full) to evalu-ate its performance on partially preprocessed short text in Serbian. We also compared it to LDA and NMF on fully preprocessed text. The experiments were conducted on a dataset of tweets expressing hesitancy toward COVID-19 vaccination. Our results show that with adequate parameter setting, BERTopic can yield informative topics even when applied to partially pre-processed short text. When the same parameters are applied in both prepro-cessing scenarios, the performance drop on partially preprocessed text is minimal. Compared to LDA and NMF, judging by the keywords, BERTopic offers more informative topics and gives novel insights when the number of topics is not limited. The findings of this paper can be significant for re-searchers working with other morphologically rich low-resource languages and short text.

Nikola Milosevic, Johannes Müller, Nico Scherf

Reinforcement Learning (RL) agents can solve diverse tasks but often exhibit unsafe behavior. Constrained Markov Decision Processes (CMDPs) address this by enforcing safety constraints, yet existing methods either sacrifice reward maximization or allow unsafe training. We introduce Constrained Trust Region Policy Optimization (C-TRPO), which reshapes the policy space geometry to ensure trust regions contain only safe policies, guaranteeing constraint satisfaction throughout training. We analyze its theoretical properties and connections to TRPO, Natural Policy Gradient (NPG), and Constrained Policy Optimization (CPO). Experiments show that C-TRPO reduces constraint violations while maintaining competitive returns.

Nikola Milosevic, Johannes Müller, Nico Scherf

In constrained Markov decision processes, enforcing constraints during training is often thought of as decreasing the final return. Recently, it was shown that constraints can be incorporated directly into the policy geometry, yielding an optimization trajectory close to the central path of a barrier method, which does not compromise final return. Building on this idea, we introduce Central Path Proximal Policy Optimization (C3PO), a simple modification of the PPO loss that produces policy iterates, that stay close to the central path of the constrained optimization problem. Compared to existing on-policy methods, C3PO delivers improved performance with tighter constraint enforcement, suggesting that central path-guided updates offer a promising direction for constrained policy optimization.

Nikola Milosevic, Gesine Müller, Jan Huisken et al.

In this work, we introduce the concept of Active Representation Learning, a novel class of problems that intertwines exploration and representation learning within partially observable environments. We extend ideas from Active Simultaneous Localization and Mapping (active SLAM), and translate them to scientific discovery problems, exemplified by adaptive microscopy. We explore the need for a framework that derives exploration skills from representations that are in some sense actionable, aiming to enhance the efficiency and effectiveness of data collection and model building in the natural sciences.

Miloš Košprdić, Nikola Prodanović, Adela Ljajić et al.

Supervised named entity recognition (NER) in the biomedical domain depends on large sets of annotated texts with the given named entities. The creation of such datasets can be time-consuming and expensive, while extraction of new entities requires additional annotation tasks and retraining the model. To address these challenges, this paper proposes a method for zero- and few-shot NER in the biomedical domain. The method is based on transforming the task of multi-class token classification into binary token classification and pre-training on a large amount of datasets and biomedical entities, which allow the model to learn semantic relations between the given and potentially novel named entity labels. We have achieved average F1 scores of 35.44% for zero-shot NER, 50.10% for one-shot NER, 69.94% for 10-shot NER, and 79.51% for 100-shot NER on 9 diverse evaluated biomedical entities with fine-tuned PubMedBERT-based model. The results demonstrate the effectiveness of the proposed method for recognizing new biomedical entities with no or limited number of examples, outperforming previous transformer-based methods, and being comparable to GPT3-based models using models with over 1000 times fewer parameters. We make models and developed code publicly available.

Nikola Milosevic, Wolfgang Thielemann

Biomedical research is growing at such an exponential pace that scientists, researchers, and practitioners are no more able to cope with the amount of published literature in the domain. The knowledge presented in the literature needs to be systematized in such a way that claims and hypotheses can be easily found, accessed, and validated. Knowledge graphs can provide such a framework for semantic knowledge representation from literature. However, in order to build a knowledge graph, it is necessary to extract knowledge as relationships between biomedical entities and normalize both entities and relationship types. In this paper, we present and compare few rule-based and machine learning-based (Naive Bayes, Random Forests as examples of traditional machine learning methods and DistilBERT, PubMedBERT, T5 and SciFive-based models as examples of modern deep learning transformers) methods for scalable relationship extraction from biomedical literature, and for the integration into the knowledge graphs. We examine how resilient are these various methods to unbalanced and fairly small datasets. Our experiments show that transformer-based models handle well both small (due to pre-training on a large dataset) and unbalanced datasets. The best performing model was the PubMedBERT-based model fine-tuned on balanced data, with a reported F1-score of 0.92. DistilBERT-based model followed with F1-score of 0.89, performing faster and with lower resource requirements. BERT-based models performed better then T5-based generative models.

Nikola Milosevic, Gangamma Kalappa, Hesam Dadafarin et al.

Medical health records and clinical summaries contain a vast amount of important information in textual form that can help advancing research on treatments, drugs and public health. However, the majority of these information is not shared because they contain private information about patients, their families, or medical staff treating them. Regulations such as HIPPA in the US, PHIPPA in Canada and GDPR regulate the protection, processing and distribution of this information. In case this information is de-identified and personal information are replaced or redacted, they could be distributed to the research community. In this paper, we present MASK, a software package that is designed to perform the de-identification task. The software is able to perform named entity recognition using some of the state-of-the-art techniques and then mask or redact recognized entities. The user is able to select named entity recognition algorithm (currently implemented are two versions of CRF-based techniques and BiLSTM-based neural network with pre-trained GLoVe and ELMo embedding) and masking algorithm (e.g. shift dates, replace names/locations, totally redact entity).

Nikola Milosevic, Junfan Huang

In the past decade, the cyber-crime related to mobile devices has increased. Mobile devices, especially the ones running on Android operating system are particularly interesting to malware creators, as the users often keep the biggest amount of personal information on their mobile devices, such as their contacts, social media profiles, emails, and bank accounts. Both dynamic and static malware analysis is necessary to prevent and detect malware, as both techniques have their benefits and shortcomings. In this paper, we propose a deep learning technique that relies on LSTM and encoder-decoder neural network architectures for dynamic malware analysis based on CPU, memory and battery usage. The proposed system is able to detect and notify users about anomalies in system that is likely consequence of malware behaviour. The method was implemented as a part of OWASP Seraphimdroids anti-malware mechanism and notifies users about anomalies on their devices. The method proved to perform with an F1-score of 79.2%.

Maksim Belousov, Nikola Milosevic, Ghada Alfattni et al.

Monitoring the administration of drugs and adverse drug reactions are key parts of pharmacovigilance. In this paper, we explore the extraction of drug mentions and drug-related information (reason for taking a drug, route, frequency, dosage, strength, form, duration, and adverse events) from hospital discharge summaries through deep learning that relies on various representations for clinical named entity recognition. This work was officially part of the 2018 n2c2 shared task, and we use the data supplied as part of the task. We developed two deep learning architecture based on recurrent neural networks and pre-trained language models. We also explore the effect of augmenting word representations with semantic features for clinical named entity recognition. Our feature-augmented BiLSTM-CRF model performed with F1-score of 92.67% and ranked 4th for entity extraction sub-task among submitted systems to n2c2 challenge. The recurrent neural networks that use the pre-trained domain-specific word embeddings and a CRF layer for label optimization perform drug, adverse event and related entities extraction with micro-averaged F1-score of over 91%. The augmentation of word vectors with semantic features extracted using available clinical NLP toolkits can further improve the performance. Word embeddings that are pre-trained on a large unannotated corpus of relevant documents and further fine-tuned to the task perform rather well. However, the augmentation of word embeddings with semantic features can help improve the performance (primarily by boosting precision) of drug-related named entity recognition from electronic health records.

Maksim Belousov, Nikola Milosevic, William Dixon et al.

Adverse drug reactions (ADRs) are unwanted or harmful effects experienced after the administration of a certain drug or a combination of drugs, presenting a challenge for drug development and drug administration. In this paper, we present a set of taggers for extracting adverse drug reactions and related entities, including factors, severity, negations, drug class and animal. The systems used a mix of rule-based, machine learning (CRF) and deep learning (BLSTM with word2vec embeddings) methodologies in order to annotate the data. The systems were submitted to adverse drug reaction shared task, organised during Text Analytics Conference in 2017 by National Institute for Standards and Technology, archiving F1-scores of 76.00 and 75.61 respectively.

Nikola Milosevic, Dimitar Marinov, Abdullah Gok et al.

In the past decade, social innovation projects have gained the attention of policy makers, as they address important social issues in an innovative manner. A database of social innovation is an important source of information that can expand collaboration between social innovators, drive policy and serve as an important resource for research. Such a database needs to have projects described and summarized. In this paper, we propose and compare several methods (e.g. SVM-based, recurrent neural network based, ensambled) for describing projects based on the text that is available on project websites. We also address and propose a new metric for automated evaluation of summaries based on topic modelling.

Nikola Milosevic, Goran Nenadic

Simile is a figure of speech that compares two things through the use of connection words, but where comparison is not intended to be taken literally. They are often used in everyday communication, but they are also a part of linguistic cultural heritage. In this paper we present a methodology for semi-automated collection of similes from the World Wide Web using text mining and machine learning techniques. We expanded an existing corpus by collecting 442 similes from the internet and adding them to the existing corpus collected by Vuk Stefanovic Karadzic that contained 333 similes. We, also, introduce crowdsourcing to the collection of figures of speech, which helped us to build corpus containing 787 unique similes.

Nikola Milosevic, Goran Nenadic

Similes are natural language expressions used to compare unlikely things, where the comparison is not taken literally. They are often used in everyday communication and are an important part of cultural heritage. Having an up-to-date corpus of similes is challenging, as they are constantly coined and/or adapted to the contemporary times. In this paper we present a methodology for semi-automated collection of similes from the world wide web using text mining techniques. We expanded an existing corpus of traditional similes (containing 333 similes) by collecting 446 additional expressions. We, also, explore how crowdsourcing can be used to extract and curate new similes.

Nikola Milosevic

Long term investment is one of the major investment strategies. However, calculating intrinsic value of some company and evaluating shares for long term investment is not easy, since analyst have to care about a large number of financial indicators and evaluate them in a right manner. So far, little help in predicting the direction of the company value over the longer period of time has been provided from the machines. In this paper we present a machine learning aided approach to evaluate the equity's future price over the long time. Our method is able to correctly predict whether some company's value will be 10% higher or not over the period of one year in 76.5% of cases.

Nikola Milosevic

People are producing more written material then anytime in the history. The increase is so high that professionals from the various fields are no more able to cope with this amount of publications. Text mining tools can offer tools to help them and one of the tools that can aid information retrieval and information extraction is semantic text annotation. In this report we present Marvin, a text annotator written in Java, which can be used as a command line tool and as a Java library. Marvin is able to annotate text using multiple sources, including WordNet, MetaMap, DBPedia and thesauri represented as SKOS.

Nikola Milošević

In past three decades almost everything has changed in the field of malware and malware analysis. From malware created as proof of some security concept and malware created for financial gain to malware created to sabotage infrastructure. In this work we will focus on history and evolution of malware and describe most important malwares.

Nikola Milošević

In linguistic morphology and information retrieval, stemming is the process for reducing inflected (or sometimes derived) words to their stem, base or root form; generally a written word form. In this work is presented suffix stripping stemmer for Serbian language, one of the highly inflectional languages.